Pre-Processing of Multi-Channel EEG for Improved Compression

Performance using SPIHT

Hoda Daou
1
and Fabrice Labeau
1
AbstractA novel technique for Electroencephalogram
(EEG) compression is proposed in this article. This technique
makes use of the inter-channel redundancy present between
different EEG channels of the same recording and the intra-
channel redundancy between the different samples of a specic
channel. It uses Discrete Wavelet Transform (DWT) and Set
partitioning in hierarchical trees (SPIHT) in 2-D to code the
EEG channels. Smoothness transforms are added in order to
guarantee good performance of SPIHT in 2-D. Experimental
results show that this technique is able to provide low dis-
tortion values for high compression ratios (CRs). In addition,
performance results of this method do not vary a lot between
different patients which proves the stability of the method when
used with recordings of different characteristics.
I. INTRODUCTION
Electroencephalography is the monitoring or recording of
electrical activity in the brain. Recording techniques of EEG
involve placing electrodes either inside the brain, over the
cortex under the skull, or at certain locations over the scalp.
Electroencephalography provides an insight on the human
brain: it can detect abnormalities, diagnose mental disorders
like dementia, epileptic seizures and psychiatric disorders.
In addition, effects of the administered drugs on the changes
in EEG waveforms can be observed. EEG is also used in
telemedicine and brain computer interface (BCI).
Recording of EEG is done over several hours and along
several channels. The number of channels can even exceed
256 for increased accuracy and reliability in diagnosis.
Every sample of EEG might be important and cannot be
disregarded. The recording can result in huge amounts of
data to be stored and/or transmitted, which calls for efcient
and low distortion compression techniques. Although lossless
compression is more desired for medical signals, it was
shown that higher compression rates can be achieved using
lossy techniques.
Scalp recordings of EEG can be seen as the measure of
the projection of the activity inside the brain on certain
locations on the scalp. The same source of activity can
have several projections on the scalp. For this reason, when
reading different EEG channels, a lot of similarity and even
superposition of the signals can be noticed. When aiming at
compression of these EEG channels, looking at the signals in
This work was supported by the Natural Sciences and Engineering
Research Council (NSERC) and industrial and government partners, through
the Healthcare Support through Information Technology Enhancements
(hSITE) Strategic Research Network.
1
F. Labeau and H. Daou are with the Department of Electrical and
Computer Engineering at McGill University, Montreal, QC H3A 2A7,
CANADA fabrice.labeau at mcgill.ca, hoda.daou
at mail.mcgill.ca.
a spatial dimension, rather than just time dimension, can help
capture more redundancy and similarities. Therefore, better
compression can be achieved when combining all channels
used in the recording.
As mentioned previously, visual correlation is highly vis-
ible both between several channels and in a single channel
between different time segments. This correlation or redun-
dancy should be exploited when building a compression al-
gorithm. In compressing medical signals, it very important to
maintain all relevant information contained in the signals for
accurate medical diagnosis. Thus, an accurate reconstruction
of the EEG signals is a very important factor, specially in
tele-medicine where noise, jitter and delays, caused by the
telecommunication technique used to transmit the signals,
add further distortion to the compression. The suggested
method aims at exploring these correlations in order to
achieve better compression with low distortion using classic
transforms and coding techniques.
The following paragraphs provide a literature review on
the usage of Wavelet Transform (WT) and set partitioning
in hierarchical trees (SPIHT) in compressing EEG signals.
Afterwards, a novel method of applying these compression
methods on EEG channels is suggested. The article ends with
an analysis on the simulations results and suggestions for
improvement.
Cardenas-Barrera et al. use Wavelet Packet Transform
(WPT) to segment and decompose the EEG Signals [1]. The
compression algorithm is composed of the following sec-
tions: segmentation, transformation, thresholding of the low-
relevance coefcients, quantization and Run-Length Coding
(RLC). Calculating the proper thresholds is the main issue
in this model, these values should preserve the signals
characteristics while keeping the distortion within acceptable
limits. This compression algorithm is able to achieve a
Compression Ratio (CR) of 9.06 with a Percent Root-mean
squared distortion (PRD) of 5.3275. CR is dened as the ratio
of the number of bits used to represent the original data to
the number of bits required to code the compressed data.
This method does not examine the mutual information that
exists between the different channels of the same recording.
A 2-D compression technique that uses an integer lifting
wavelet transform (ILWT) as the de-correlator, with SPIHT
as the source coder is presented [2]. The 2-D algorithm is
compared to the same one in 1-D that also uses ILWT with
SPIHT and the 1-D case gave much higher distortion for the
same bit rate and larger delays. 2-D SPIHT coding is suited
for smooth natural images. However, EEG signals possess a
non-stationarity characteristic. Thus, the 2-D matrix formu-
34th Annual International Conference of the IEEE EMBS
San Diego, California USA, 28 August - 1 September, 2012
2232 978-1-4577-1787-1/12/$26.00 2012 IEEE
lation of the channels is not as smooth as natural images
and the compression performance deteriorates for certain
segments. Smoothness transforms should be added to the 2-D
EEG matrix to enhance the performance of SPIHT. Authors
do not show how distortion varies at different CRs, however,
bit planes are used to display the results of the compression.
EEG Compression in 3-D was applied by Dauwels et
al. and tested on two types of data [3]. Methods like
SVD (Singular Value Decomposition), Parafac and Tucker
decompositions were tested on recordings of patients with
mild cognitive impairment (MCI) and on EEG-Motor Men-
tal Imagery datasets of the physiobank database [4]. Rate
distortion curves show that Tucker and Parafac gave the best
results in compression. However results vary a lot between
the two different types of recording. Better results were found
for the second dataset where the subject opens and closes
his st until a certain target disappears from a screen. When
reading this data, peaks are observed every few seconds and
on all channels. This data is very articial and does not
reect the characteristics of EEG data types that are usually
recorded over long periods of time like long-term monitoring
of patients with epilepsy.
A recent article published in Electronic Letters studies the
compression of EEG using 1-D SPIHT [5]. In this article, the
computational complexity of the method is analysed. Their
results show that 1-D SPIHT is able to achieve a CR of about
30 for a PRD of 30%. Testing was done on intra-cranial
recordings of 21 patients suffering from epileptic seizures.
However, when compressing EEG signals, results can vary
a lot between different types of recordings. 1-D SPIHT is
tested on scalp recordings in this article and the results are
shown in section IV. Authors of this article do not show
how the results are varying between the different patients.
In addition, this algorithm compresses each channel alone
and does not take into account the inter-channel redundancy
that exists between the different channels. This redundancy is
more present in intra-cranial electrodes that provide record-
ings at the same location but at different depths.
This article presents a lossy EEG compression scheme
that applies image compression techniques on 2-D EEG
matrices. Spatial correlation between different EEG channels
is reduced using 2-D DWT and SPIHT. A method that adds
smoothness to the EEG matrix is suggested. This is done
to enhance the performance of SPIHT and achieve better
performance in compression.
II. METHODS
A. Discrete Wavelet Transform (DWT)
Wavelet Transforms provide multi-resolution, locality, and
compression when combined with zero-tree coding tech-
niques. Many compression algorithms use these transforms
to decompose a signal and take advantage of the properties
of these coefcients in energy compaction.
DWT gives the time-scale representation of a digital
signal using digital ltering techniques. To nd the DWT
coefcients, the signal is passed through cascades of low and
high pass lters implemented at the low frequency bands of
each level, the high frequency bands are left unchanged. The
resolution of the signal, which is a measure of the amount of
detail information in the signal, is changed by the ltering
operations, and the scale is changed by up-sampling and
down-sampling operations.
B. Set partitioning in hierarchical trees (SPIHT)
The Set-Partitioning in Hierarchical Trees (SPIHT) is a
coding algorithm that exploits the relationships between the
wavelet coefcients across the different scales at the same
spatial location in the wavelet sub-bands [6], [7]. Exact bit
usage control can be achieved using the SPIHT algorithm.
A pre-specied bit-rate or quality requirement can be used
as criterion to stop the encoding and decoding process at
any instance [8]. SPIHT targets the coding of the position of
signicant wavelet coefcients and the coding of the position
of zero-trees in the wavelet sub-bands. It was originally
suggested for the compression of 2-D images, thus it exploits
the basic characteristics of this type of data. More precisely,
it exploits the following image characteristics [6]:
1) Most of the images energy is located in the low fre-
quency components and there is a decrease in variance
as we move from the highest to the lowest levels of
the frequency sub-band pyramid.
2) Spatial self-similarity is observed among the sub-
bands, and the coefcients are likely to be better
magnitude-ordered when going deeper in the frequency
sub-band pyramid along the same spatial orientation.
Signals smoothness can be measured by the amount of
energy in the low frequency bands [9]. When working with 2-
D matrices, one level of wavelet decomposition produces an
approximate sub-band (low frequency (LF) band) and three
high frequency (HF) bands. The matrix is considered smooth
when more energy is concentrated in the LF band compared
to HF bands.
III. COMPRESSION SCHEME
SPIHT coding originally targets two dimensional images.
In EEG, transferring the data into two-dimensional matrices
can be accomplished either in one channel or in a multi-
channel context. However, choosing the appropriate segment
and matrix size and the optimal channel numbering in the
multi-channel case can highly affect the performance of the
compression.
The 2-D multi-channel matrix is formed by choosing the
rows of the matrix as segments from different channels as
follows:
EEG
i
=

s
i,1
s
i,2
.
.
.
s
i,M

(1)
where s
i,l
is the EEG segment of index i of channel at index
1 and M is equal to the number of channels used in the
recording. Each EEG segment used in the matrix groups N
consecutive time samples. Arrangements of the channels in
the matrix affects the smoothness of the matrix and thus the
performance of the compression.
2233
The algorithm rst computes the correlation coefcient of
each row of the matrix with all other rows. To build the
smooth EEG matrix, the rst row is kept as the rst EEG
channel used in the recording, then each adjacent row is
chosen as the highest correlated row among all other rows
in the matrix. When doing this, in most matrices, the last
rows are the channels that are close to the reference. This is
due to the fact that there is least amount of activity in this
region.
This method of choosing the rows of the matrix guarantees
that each two adjacent rows are highly correlated. To achieve
more smoothness, the rst P+1 rows are selected. The value
P is chosen as the highest value that is smaller than M
and divisible by 2
k
, with k equal to the number of levels
used in the DWT. From these P+1 rows or channels, the
channel that has the highest correlation coefcient with all
other P channels, Re f , is selected. The values of its samples
are subtracted from each channel of the P-by-N matrix.
This results in added smoothness to the 2-D EEG matrix.
DWT is then applied on this P-by-N matrix. However,
high correlation is still present within the blocks of the
DWT coefcients of the same sub-band, which is common
when compressing naturally smooth images [10]. To reduce
this, a de-correlation block, that involves simply applying
Discrete Cosine Transform (DCT), was added after DWT
[10]. SPIHT coding is then performed on the transformed
DWT coefcients.
The reference channel and the other least correlated chan-
nels are coded using 1-D DWT and SPIHT. The indices of
the channels showing the chosen arrangement in the matrix,
and the index of the highest correlated row, Re f , are sent as
overhead with SPIHT output. This coding scheme is repeated
for every block of M N samples in the recording.
IV. RESULTS AND DISCUSSIONS
In the scalp recordings done at the EEG lab at the Montreal
Neurological Institute (MNI) that are used in this paper, 29
electrodes are used with a sampling frequency of 200 Hz.
The montage used is referential with the reference located
on the center of the scalp, known as FC
z
in the International
10-20 System [11, p.139].
The performance parameter used to analyse the results is
the percent-root mean square distortion (PRD):
PRD
i,l
(%) =

N
n
(s
i,l
[n] s
i,l
[n])
2

N
n
s
i,l
[n]
2
100 (2)
where s
i,l
[n] is the EEG sample n of segment at index i and
channel at index l, and s
i,l
is the reconstructed EEG segment
after compression. Thus PRD is calculated for each EEG
channel, l, and segment i, then the mean over all segments
and channels is calculated to reect the PRD at a certain CR.
In the testing, M is equal to 29, resulting in P equal to
24 for a DWT number of levels, k, equal to 3. Values of k
equal to 2, 3 and 4 were tested and 3 was chosen because
it guarantees that the least correlated channels coded using
1-D SPIHT are the four channels closest to the reference. It
gave better performance than the other values. For choosing
the appropriate wavelet in the DWT, it is suggested that
Biorthogonal 4.4 wavelet has a more established popularity
in compressing 2-D images [12], [13]. The compression
method re-arranges the 2-D matrix in order to achieve more
smoothness. The transformed EEG matrices have almost
similar properties as natural images. For this reason, in our
testing, biorthogonal 4.4 proved to be better than the classic
Haar wavelet.
Fig. 1 shows the PRD results for different compression
ratios of 1-D SPIHT and 2-D SPIHT. 1-D SPIHT was used
with segment length N equal to 1024, which gave the best
results compared to other values like 128, 256 and 512. The
size of the EEG channel segments used in the 2-D SPIHT
method is equal to 256. A 5 level DWT with biorthogonal
4.4 wavelet, was performed prior to SPIHT coding in the
1-D case. This plot highlights the fact that 2-D SPIHT is
able to achieve higher compression ratios for the same PRD
percentage. This is due to the fact that in the one dimensional
case, the correlation between the channels still exists after
compression. Thus, redundant information is still being sent.
In addition, when applying SPIHT in 2-D, the entropy of the
output is close to 0.5. Whereas, for 1-D SPIHT, the entropy is
close to 1. For this reason, a Run-Length coding (RLC) block
was added to 2-D SPIHT output and there was a relevant
decrease in the total number of bits required for coding a
specic matrix.
1-D SPIHT is shown to achieve a CR of about 30 for
a PRD of 30% when tested on intra-cranial recordings of
patients with epilepsy [5]. However, when testing on the
scalp recordings obtained from MNI, results are different and
the proposed method is able to achieve better compression
performance. This highlights the fact that EEG recordings
differ a lot between different types of recording methods
and different patients.
The proposed method is compared to Tucker and Parafac
tensor decompositions applied on 3-D EEG tensors [3]. The
size of the tensor used in these two methods is equal to
29-by-16-by-16. Thus the same EEG segment used in 2-D
SPIHT of size N equal to 256 is arranged in 2-D as 16-
by-16 as explained in the article [3]. The third dimension
in these two methods is the EEG channel number. Testing
was done on 9 patients over a period of one hour for
each patient. Fig. 3 shows the PRD results with the 25
and 75 percentiles to highlight how the PRD values are
varying between the patients. The proposed method gave
better performance results than Tucker and Parafac. However
for CRs higher than 18.5, the performance of the new method
and Parafac are almost the same. A 7% PRD is suggested
to be the maximum allowed loss for preserving clinically
relevant information [1], [5]. For this PRD value, Parafac is
able to achieve a CR of around 2, while the proposed method
can achieve a CR of around 5. The methods were also tested
for N equal to 1, 024, resulting in 29-by-32-by-32 tensors,
but the same improvement in performance was noticed. The
only difference in the results is all three PRD lines drop by
almost 3%.
For compression ratios below 20, the PRD results vary
2234
2 4 6 8 10 12 14 16 18 20
5
10
15
20
25
30
35
40
45
Compression Ratio
P
R
D

(
%
)
1D SPIHT
2D SPIHT
Fig. 1. Percent-root mean square distortion comparison between 2-D and
1-D SPIHT.
much less between patients for the proposed method than
for Parafac and Tucker. This variance is negligible for CRs
below 7. This criterion is very important in EEG compression
since the characteristics of the recordings vary a lot between
different patients.
V. CONCLUSIONS AND FUTURE WORKS
A. Conclusions
Looking at the EEG channels in 2-D enables us to make
use of the redundancy found between the channels and
between the samples of the same channel. The proposed al-
gorithm is able to achieve low variance between patients and
low distortion compared to other compression methods like
1-D SPIHT, Tucker and Parafac. For PRD values lower than
30%, the proposed algorithm achieves higher compression
ratios. Above this value, distortion can be regarded as high
which is to be avoided when dealing with biomedical signals
with important diagnostic information.
B. Future Works
EEG recordings vary a lot between both different types
of recordings and different patients of the same type of
recordings. Thus, nding an algorithm that has stable perfor-
mance is a challenge. In addition, the compression scheme
should still preserve important diagnosis information. It
would be important to test abnormality detection systems,
like epileptic seizure detection, on both the original data and
the compressed output to further analyse the performance
of the compression algorithm. In addition, the performance
of the compression should be tested on different types of
recordings like invasive intra-cranial EEGs.
ACKNOWLEDGMENT
The authors gratefully acknowledge Professor Jean Got-
man and his team at the Montreal Neurological Institute
and Hospital, McGill University, for providing the long-term
EEG recordings.
2 4 6 8 10 12 14 16 18 20
5
10
15
20
25
30
35
40
45
Compression Ratio
P
R
D

(
%
)
Tucker
Parafac
2D SPIHT
Fig. 2. Percent-root mean square distortion comparison between Tucker
[3], Parafac [3] and the proposed method. The boxes extend from the 25th
percentile to the 75th percentile.
REFERENCES
[1] J. Cardenas-Barrera, J. Lorenzo-Ginori, and E. Rodriguez-Valdivia,
A wavelet-packets based algorithm for EEG signal compression,
Informatics for Health and Social Care, vol. 29, no. 1, pp. 1527,
2004.
[2] K. Srinivasan and M. Reddy, Efcient preprocessing technique
for real-time lossless EEG compression, Electronics Letters,
vol. 46, no. 1, pp. 2627, 2010. [Online]. Available:
http://link.aip.org/link/?ELL/46/26/1
[3] J. Dauwels, K. Srinivasan, M. R. Reddy, and A. Cichocki, Multi-
channel EEG compression based on matrix and tensor decomposi-
tions, in Proceedings of the 2011 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), July 2011, pp.
629 632.
[4] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C.
Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, , and H. E.
Stanley, Physiobank, physiotoolkit, and physionet: Components of a
new research resource for complex physiologic signals, Circulation,
vol. 101, no. 23, p. 215

A S 220, 2000.
[5] G. Higgins, B. McGinley, N. Walsh, M. Glavin, and E. Jones,
Lossy compression of EEG signals using SPIHT, Electronics Letters,
vol. 47, no. 18, September 2011.
[6] D. Rawat, C. Singh, and M. Sukadev, A hybrid coding scheme
combining SPIHT and SOFM based vector quantization for effectual
image compression, European Journal of Scientic Research, 2009.
[7] Z. Lu, Y. Kim, Z. Lu, D. Y. Kim, and W. Pearlman, Wavelet
compression of ECG signals by the set partitioning in hierarchical trees
(SPIHT) algorithm, IEEE Transactions on Biomedical Engineering,
vol. 47, pp. 849 856, 1999.
[8] D. Y. Kim, Z. Lu, and W. A. Pearlman, Wavelet Compression of ECG
Signals by the Set Partitioning in Hierarchical Trees Algorithm, IEEE
Transactions on Biomedical Engineering, 2000.
[9] K. Srinivasana, J. Dauwels, and M. Reddy, A two-dimensional ap-
proach for lossless EEG compression, Biomedical Signal Processing
and Control, 2011.
[10] K.-L. Kim and S.-W. Ra, Performance improvement of the SPIHT
coder, Sig. Proc.: Image Comm., vol. 19, no. 1, pp. 2936, 2004.
[11] E. Niedermeyer and F. Da Silva, Eds., Electroencephalography, 5th ed.
Lippincott Williams and Wilkins, 2005, vol. 7.
[12] J. Azpiroz-Leehan and J.-F. Lerallut, Selection of biorthogonal
lters for image compression of MR images using
wavelet packets, Medical Engineering and Physics, vol. 22,
no. 5, pp. 335 343, 2000. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1350453300000424
[13] V. K. Bairagi and A. M. Sapkal, Selection of wavelets for medical
image compression, in International Conference on Advances in
Computing, Control, and Telecommunication Technologies, 2009, pp.
678 680.
2235